How SenseNova U1.5 Unifies Visual Understanding and 4K Image Generation

SenseNova U1.5 brings image understanding, visual reasoning, generation, and editing into an 8B-parameter Mixture-of-Transformers (MoT) model. Rather than handing images to an external vision encoder and decoding generated images through a VAE, it uses a shared visual interface and reconstructs output in RGB pixel space. SenseTime’s technical report presents that design as a way to retain visual understanding while improving image creation at resolutions up to 4K. 13

How can one model read and create images?

U1.5 represents image regions as 32×32-pixel visual tokens. Those tokens give the model a compact sequence for visual tasks without a separate vision encoder. On the generation side, its decoder turns the model’s visual states into RGB pixels rather than relying on a VAE to decode an image latent. That shared interface is the basis for combining perception, reasoning, and creation in one architecture. 121

The consequential change from U1 is how those pixels are reconstructed. U1 used a patch-wise multilayer perceptron (MLP) head that predicted each image patch independently, which could expose boundaries as seams or grid artifacts at high resolution. U1.5 places the visual states back into a two-dimensional grid and uses a lightweight spatial decoder with progressive Pixel Shuffle and 3×3 convolutions. Neighboring regions can therefore influence reconstruction instead of being rendered in isolation. 1322

For generation up to 4096×4096, SenseTime also describes resolution-aware noise conditioning: a resolution-dependent noise-scale embedding is combined with the generation timestep so the model can account for changes in noise behavior as output size changes. The spatial decoder addresses continuity across patches; the noise conditioning addresses the change in scale. Together, they are intended to make native 4K generation more stable, not to guarantee artifact-free results. 1329

Why train specialists if the result is one model?

SenseTime’s post-training approach is to specialize, then unify. It optimizes experts for visual aesthetics, Chinese and English text rendering, infographics, and image editing, then consolidates their capabilities through multi-expert on-policy distillation. The separate experts are part of the training process; the reported outcome is a unified model, not a requirement to run four specialist models for every request. 129

Compared with U1, SenseTime reports stronger high-resolution detail and spatial continuity, along with improvements in text rendering, layouts, and editing, while retaining the encoder-free, VAE-free understanding-and-generation design. Reported U1.5 scores include 0.92 on GenEval, 0.948 on CVTG-2K, and 4.59 on ImgEdit. These are the reported evaluation results, not independent proof of an improvement on every visual task. 1328

What has been released?

The U1.5 technical report is public, and an official Hugging Face page identifies a SenseNova-U1.5-8B-MoT checkpoint. That full-model listing is distinct from the U1.5-8B-MoT-Preview page and SenseTime’s separately announced U1.5-Lite-Preview release; the preview assets should not be treated as the full checkpoint. 1312213

SenseTime says it will open-source training code for supervised fine-tuning, reinforcement learning, and on-policy distillation. The report and model listings do not, by themselves, establish that the complete promised training-code release is already available. 13